Lyron
User
Newbie
| Posts: 6 |   |
Karma: 0
|
Fritz 8 - Questions about learning - 2007/01/24 13:11
I am a beginner and have been playing "Rated" games agst Fritz8, i.e where you cant take back moves. It helps me keep a good track of how i am doing, and i can always go back and try a different line when the game is over, it even remembers the times!
When i first started playing rated i set the "playing strength" slider all the way to the left. It was 1540 originaly but after a few games the minimum value seems to be higher (1590). All this without transfering the book on the hard disk, i.e no learning. Obviously the computer must somehow keep track of how well i am doing and incresed its minimum strength.
So my first question is what does the strength relate to, in laymans terms?
It cant be associated with the tree, thats different. It must have something to do with when the engine goes 'out of book' So how does it do it? does it literarly slow itself down by counting from 0 to a gazillion? does it adjust the search depth? Anyway...lets move on
So i have played ard 200 games, all kept in the autosave database. I have found a couple of winning "sequences" which help me keep my win-loss averages between 1-2 and 1-3. And this is where i am troubled. The whole point is to learn how to play. So if i find a winning sequence its almost like cheating. I do think there is value in going through the sequence again because i remember it better, and i expect that its the same in real life, as you can beat different people the same way as well as make the same mistakes again. But obviously, after 10 identical games, further repetitions will make me overconfident, ignoring different lines of play. This got me to think about transfering the opening book to the hd and enabling the learning function.
Having read the manual i realise that i have several options and ...questions. First of all what is the tournament option. Are there moves that have been excluded from tournaments? What are they?
An intersting option is that it looks like i can adjust the weights based on the games that i have played so far. That is a good idea although i am worried that the weights are going to be overadjusted because of the repetitive wining sequences, which the engine would not have gone through if it was actual learning while playing each of the 200 games sequentialy (as opposed to learning after 200 games played with no adjustment, and then learning). I suppose i can set a low value on the weights for this and adjust it after its done reading the database. But i am confused about the weights too.
It seems there are three parameters (1) variety of play (2) influence of learn value (3) learning strength
Having toyed with AI in college (perhaps this is where my troubles with this come from) i would imagine that the adjustment that goes on, looks something like
new prob = old prob + influence * strength
Several questions here. First it seems that influence and strength are interchangable, not only from the equation above (which is probably oversimplified to wrong) but also from how the are explained in the manual.
Also, variety is also a bit of a mystery. It seems that what it does, is scew the distribution of moves, i.e greatest variety = uniform distribution.
Although i guess this defeates the whole learning process since uniform distribution means all moves have equal probability, it might be used for revising. I.e. by setting variety towards the maximum, the engine is forced to play lines for which it has a losing record more often, thus testing the players memory!?
Another question is how do the adjustments backpropagate, i.e if the engine loses game does the whole line of play get a "bad mark"? Presumably if the last move of a game was a blunder, this does not necessarily mean that the previous moves where also bad.
So maybe the learning strenth is constant throughout the tree but the influence varies according to how high up we are in the tree. After all if playing white, the engine loses a game it has opened with e4, this does not mean that e4 is a bad move. So maybe the adjustment looks like
new prob = old prob + F(influence) * strength
where F(x) is a function which is tree (depth) related, for example
f=(move no./total moves in the game)*influence
so that the 1st move gets (1/total moves)*influence of an adjustment, which is a lot less than the checkmate move which gets (total moves/total moves)*influence, a bigger adjustment.
Does this make sense?
So how do different settings affect the engine and more importantly the player? I am really hoping that someone has played around with these and can give some insight, or know outright.
A low strength value would mean that it would take a lot of games for the engine to make up its mind about anything, be it a move, or a line of play.
A high strength would mean that the engine would not try again a move which ended in a loss.
A low influence would mean that the engine would still play moves that later ended up in a loss, whereas high influence runs the risk of discarding whole trees.
Still this is not that clear. After all, what the engine does during play, is search. So do these values also affect which moves it searches? Possibly.
And what does increasing the playing strength in rated games do in relation to the weights?
If anyone knows or has thought about any of these....fire away.
---------
In baseball, you don't know nothin'.
Popular posts by Lyron
U.S.A. Chess Statistics - part 2
LAND OF THE *FREE*
raplacement scheme for transp. t...
|
post new topic